Tara Seshan on OpenAI's AI Product Strategy

There is an old assumption about building things: put money in, get value out, the way a real estate developer buys land and collects rent. Tara Seshan, who leads product for ChatGPT Work and Codex at OpenAI, thinks software has never really worked that way, and it works even less that way now.

"Software is not like real estate. You don't, like, put money in and get value out. It is a little bit more like filmmaking, where you can put a lot of money into a film, but that doesn't guarantee that the film is successful or good." (Patrick Collison or John Collison, as quoted by Tara Seshan, 14:23)

Seshan has spent her career testing that idea. She was one of the first five product managers at payments company Stripe, later led product at climate-accounting startup Watershed, and spent time as an entrepreneur in residence at venture firm Sutter Hill Ventures before joining OpenAI about a year ago (02:24), a stretch of time she calls a lifetime by AI standards. She now leads product for both ChatGPT Work and Codex, OpenAI's coding agent.

Her framework for where the industry is headed is simple enough to fit on a napkin. The first era of AI products was chat: type a question, get an answer. The second era is agents, software that can carry out multistep tasks with limited supervision, most visibly in coding tools. The third era, which she thinks is arriving soon, is the persistent coworker (00:00).

"That third era that might come soon is how do you work with a persistent coworker who is able to get things done with you?" (Tara Seshan, 00:00)

What struck her most about OpenAI itself was not secrecy but the opposite. She expected a hidden strategy document, a master plan. Instead she found that almost every internal idea about how the model should behave becomes public product or public messaging within weeks (04:40). The company, she says, is founder-led in an unusual sense: nearly everyone functions as a founder of their own area, with very little top-down direction (03:12).

The Two-to-Three-Month Window

That founder-like autonomy comes with a discipline that took Seshan time to adjust to. At past companies, rigorous long-range planning was rewarded: write the strategy memo, reason from first principles, predict the market's next moves. At OpenAI, she says, being prolific and empirical matters more than being academic (07:41). Rather than a long reasoning document, the job is to find the single sharpest hypothesis and test it with users as fast as possible (08:58).

The clearest expression of that discipline is a rule she repeats often: build for where the underlying AI model will be two to three months from now, not for its current limits and not for a guess about where it will be in a year (27:39). Both extremes, she argues, produce a product that breaks the moment the model changes underneath it.

"You fail if you build for where the models are now. You fail if you build for where you think the models will be in a year. Both outcomes are equally wrong." (Tara Seshan, 27:39)

Two internal phrases keep that rule alive day to day. The first is whether a project is "maximally accelerated," meaning the team is moving as fast as the opportunity allows. The second, credited partly to her engineering manager Andrew Ambrosino, is whether people are "mainlining" the product: using it all day, every day, to get their own work done, so that friction surfaces before launch rather than after (25:39). It is dogfooding with higher stakes, since the product in question is often still deciding what it wants to be.

That urgency shows up in how ChatGPT is now organized around a toggle between "chat" mode and "work" mode, with Codex sitting alongside both (29:36). Under the hood, work mode runs on the same technology as Codex, just with a friendlier interface for people who do not think of themselves as developers (30:59). Seshan's stated goal is for that toggle to disappear entirely, so a user simply describes a task and the system picks the right approach (29:56). Getting there has meant accepting an old tradeoff in reverse: in the past, she says, polish was king, because shipping early made little difference to the outcome. Now, getting something transformative into people's hands early beats waiting for it to be perfect (36:21).

Writing as Thinking

If the two-to-three-month rule governs what OpenAI builds, a quieter rule governs how Seshan works. She separates her writing into two categories. Writing as reporting covers status updates and routine summaries, work she happily hands to an AI model. Writing as thinking is different: drafting a brief about why a team should pursue a certain strategy, then cutting and revising it until the idea is sharp. That kind of writing, she says, she will never automate (53:20).

"Writing as thinking is something I never will automate. You shouldn't replace your thinking with it." (Tara Seshan, 53:20)

A related habit came from a former manager: take a document to about 70 percent completion, then bring it to the people whose support is needed, and finish the last third together (57:40). A polished, complete draft, in her experience, invites less real feedback, because good collaborators want rough edges they can still shape.

The same instinct toward ambition shows up in how she thinks about the future of knowledge work. As AI agents take on more of the routine execution, she expects work to feel more like steering a ship than rowing it, with people spending their time on direction and judgment rather than on the mechanics of getting there (11:15). Part of the product manager's job, in her view, is to push people's sense of what is possible, sometimes by simply asking whether a plan could be ten times bigger or done in a fraction of the time (24:18, 24:37).

Her time at Sutter Hill Ventures, the venture firm behind companies including cloud-data company Snowflake, taught her a related lesson about ambition applied to marketing rather than product. She had assumed product-market fit was the hard problem and positioning was secondary. She now believes the opposite is often true in business software: the narrative, tested by pitching it to roughly a hundred people, can and should be refined before the product's final shape is locked in (63:12, 63:36).

"Product market fit is sure important, but actually I really underrated product marketing fit. The idea that the way you talk about the product and the way you market it can precede actually even building the product." (Tara Seshan, 63:12)

Toward the end of the conversation, Seshan draws a distinction that may outlast most of the specific tools discussed: coding AI and knowledge-work AI are not the same problem. A piece of generated code can be checked against automated tests; it either runs or it doesn't. A slide deck or a financial analysis cannot be judged that way. There is no test to run against a strategy memo, so the process that produced it, the citations, the visible reasoning, the steps along the way, has to be part of what gets checked (65:25).

That distinction is, in a way, a restatement of her filmmaking line from earlier. A budget does not guarantee a good film, and a working test suite does not guarantee a trustworthy answer. Somewhere in between the model's growing capability and the discipline of testing a sharp hypothesis, there is still a person deciding what the work should say, and whether it is any good. Seshan's bet, quietly repeated across the whole conversation, is that this part of the job will not get automated, even as almost everything around it does.